Three DJI Tello drones guard a spot in a room against two attacking drones. A planner decides which defender covers which attacker, and a small trained neural network (our "policy") flies each defender there.
Result: in simulation, attackers are kept out in 61.3% of rounds (38.7% breach rate).
Watch it: reports/videos/ (what the videos show).
On real drones it did not work yet: see reports/4_real_drones.md and reports/videos/realtesting.mp4.
| Folder | What's inside |
|---|---|
defend_task/ |
Our simulator task: attackers, planner, rewards |
training/ |
Train the policy and score it |
real_drones/ |
Fly the trained policy on real Tellos |
policy/ |
Trained policies: shared_stage3_best.pt (submitted) and scratch_attr_s1.pt (newer, see reports/5) |
tests/ |
Quick checks (no simulator needed) |
reports/ |
Plain-English write-ups, simulation videos and the full PDF |
archive/ |
Old notes and the full git history |
organiser_base/ |
The organisers' starter kit, unchanged |
Quick checks (any laptop):
pip3 install torch numpy pyyaml scipy pytest
python3 -m pytest tests real_drones/tests -qScore the policy in simulation (NVIDIA GPU, Isaac Lab installed with
organiser_base/simulation_infra/setup.sh ~/isaac):
source ~/isaac/env_isaaclab/bin/activate
export OMNI_KIT_ACCEPT_EULA=YES TELLO_USD_PATH=$HOME/isaac/assets/tello.usd
python training/evaluate_defend.py --task AS-Defend-Role-v0 --headless --num_envs 256 \
--episodes 512 --seed 1234 --shared --checkpoint policy/shared_stage3_best.pt --env_set stage=3If you installed Isaac Lab before the folders were reorganised, first run this once:
uv pip install -e organiser_base/simulation_infra/source/action_space_kit
Fly real drones (start the organisers' drone_gui first and take off from it):
pip3 install -r real_drones/requirements.txt
python3 real_drones/run_defender.py --defenders TELLO-1,TELLO-2,TELLO-3 \
--attackers TELLO-7,TELLO-8 --policy policy/shared_stage3_best.pt --dry-runRemove --dry-run to actually send commands. Press Ctrl+C to stop; the drones hover.
How to train it yourself: reports/3_training.md.